Speaker: Manuel Baltieri (RIKEN CBS)
Title: Variational inference in agents, with connections to control theory and cognitive (neuro)science
Abstract:
Approximate Bayesian inference (ABI) methods in machine learning are now an established set of tools delivering results for problems where exact Bayesian formulations are unfeasible. In other fields, such as neuroscience, ABI techniques have been instrumental in modelling and analysing complex datasets, for example in neuroimaging.
In what is sometimes defined as the “Bayesian turn” of cognitive science, brains, and perhaps agents more in general are now thought to be performing (approximate) Bayesian inference on partially observable environmental states. Not only are Bayesian methods useful to scientists to model complex datasets, it is now believed that cognitive agents themselves are performing some kind of (approximate) Bayesian computation.
This has generated an array of different proposals attempting to describe and understand processes in computational and cognitive neuroscience as processes of ABI. Such proposals include for instance, Friston’s popular “active inference” framework, suggesting a neural code based on fixed-form (Gaussian) variational approximations, and the “sampling hypothesis” relying on MCMC sampling methods to describe the activity of neural populations.
In this presentation I will build on approaches relating to the former, however providing a few important twists including:
• a focus on behaviour and agents over neural activity and neural codes,
• connections to control theory methods in continuous-time, and building on this,
• a discussion on the duality of inference and control, and dual effects of control, in the context of cognitive and computational neuroscience.
